1. Field of the Invention
The present invention generally relates to the synchronization of multiple copies of data in a computer system. More particularly, the present invention relates to the synchronization of copies of data stored on several computer systems linked together in a cluster.
2. Description of the Related Art
In a computer program that accesses various forms of data stored on disk or other non-volatile memory, it is often advantageous to store copies of high use data in volatile memory. Accessing data from volatile memory is typically much faster than accessing data from non-volatile memory. The process of storing high use data in volatile memory is commonly called caching.
Large-scale computer applications such as transactional applications used in the financial industry often make use of caching for performance reasons. As used herein, a “transaction” is a series of instructions executed by a computer system for carrying out a financial operation. A transaction may include multiple steps. Examples of transactions include, but are not limited to, financial transactions such as deposits, withdrawals, and funds transfers between accounts. These large-scale applications have increasingly demanded more processing power. Early mainframe computer systems typically contained one processor, one volatile memory, and arrays of non-volatile memory such as disk and tape. Later, multiple-processor systems were introduced which provide a significant performance boost over single-processor systems.
One type of multiple-processor system is a Symmetric Multiprocessor system (SMP). An SMP comprises a set of processors that shares a number of common components, including volatile memory. SMPs offer significant performance advantages over single-CPU systems, and SMPs are scalable by adding additional processors. A major disadvantage of SMPs, however, is that significant resources must be dedicated to synchronizing the activities of the processors. A point may be reached at which adding additional processors actually decreases overall performance. Furthermore, in addition to the processing and scalability limitations, multiple-processor systems share a weakness with single-processor systems: when the system experiences planned or unplanned downtime, the processing services provided by the system are unavailable.
The demand for processing performance and scalability greater than that provided by multiple-processor systems and the demand for increased availability led to the development of clusters. In general, a cluster is a group of servers that can share resources and cooperate in processing.
Clusters may be categorized into three general types. The first type is a high-availability cluster. In a high-availability cluster, one server is typically the primary server and a second server is typically the backup server. If the primary server becomes unavailable due to failure or other reasons, processing is transferred to the backup server. In order to provide as seamless a transition as possible, the backup server is commonly kept in an idle state, negatively impacting the price-performance ratio of the system. High-availability clusters increase the availability of computing resources, but they do not provide significantly greater performance than earlier systems.
A second cluster type is the parallel cluster. Parallel clusters include two or more servers that share some resources; work may be split among the servers and performed in parallel. The servers in the cluster appear as multiple logical systems to the client. During processing, programs and data may be partitioned among the servers. The partitioning of program workload and data among the servers is difficult to manage. No single server has access to all of the data. Adding additional servers to the cluster requires repartitioning of the workload and data. Unbalanced workloads or data may overload one or more servers while other servers are underutilized. If a server fails, its workload must be passed to a backup server to maintain availability. This can be detrimental to performance if the backup server has a workload of its own or has a slow access path to the failed server's data.
The third cluster type is the single-system image cluster. The servers in a single-system image cluster appear as one logical system to clients and to application programs mnning on the cluster, hence the name “single-system.” All servers in the cluster have access to all of the data, and any server in the cluster may run any transaction. Copies of a program may run concurrently on several servers in the cluster. The workload may be dynamically distributed among the servers. The single-system image cluster solves the availability and scalability problems and adds a level of stability by the use of redundant systems with no single points of failure. Effectively, the one logical system may be available year-round to clients and application programs without any outages. Hardware and software maintenance and upgrades may be performed without the loss of availability of the cluster and with little or no impact to active programs. The combination of availability, scalability, processing capability, and the logical system image make the single-system image cluster a powerful environment on which to base a large-scale transaction-based enterprise server.
Single-system image clusters typically share external non-volatile data storage, such as disk drives. Databases and other types of data permanently reside on the external storage. The servers, however, do not generally share volatile memory. Each server in the cluster operates in a dedicated local memory space.
As mentioned above, copies of a program may reside on several servers, and the programs may appear as one program to the client. Workload is distributed among the copies of the program. The programs may access shared data from external storage. For performance advantages, each program may cache high-use data in local memory space.
A single-system image cluster may include at least one Coupling Facility (CF) which provides hardware and software support for the cluster's data sharing functions. Typically, a CF may provide data synchronization for proprietary database architectures such as VSAM, IMS, and DB2, but a CF usually does not provide an application programming interface (API) to allow third-party application programs to directly access the data synchronization functions of the CF. Because of the lack of an API, the CF's data synchronization capabilities cannot be easily applied to data architectures other than the a handful of proprietary architectures listed above.
The single-system image cluster may also provide a timer facility to maintain time synchronization among the servers. On a single-system image cluster, several operating system images such as MVS images may be running on at least one computer system. MVS and OS/390 are examples of mainframe operating systems. OS/390 is a newer version of the MVS operating system, and the terms OS/390 and MVS are used interchangeably herein. “MVS image” is used synonymously with “server” herein. Operating systems other than MVS may also run as servers on a single-system image cluster. Each server is allocated its own local memory space. The servers appear as one logical server to a client. Programs may be duplicated in the memory space of several servers. The workload of a program may be divided among several copies of the program running on different servers. The data required by the program is stored on shared non-volatile memory. A copy of the program may cache frequently used portions of the shared data in a local memory space. As in the case with the multiple servers appearing as one logical server, multiple copies of a program running on a single-system image cluster may appear as one logical program to the client.
The caching of copies of data in local memory areas by programs running on servers in a cluster may lead to a problem in a single-system image cluster environment. If copies of a program require the same data, each may cache a copy in local memory. If one copy of the program modifies a local copy of the data and, optionally, the base data on the external storage, copies of the data cached on other servers are no longer concurrent with the modified copy. It is therefore desirable to provide a method of efficient, timely, and reliable synchronization of data in a single-system image cluster environment. It is also desirable to provide an API to the coupling facility to allow software developers to make source code calls to the functions of the coupling facility, and to enforce data synchronization for data architectures other than proprietary datbases.
The problem of data synchronization may also occur in computer systems in general where a program or programs make copies of shared data. Therefore, a solution to the data synchronization problem should preferably be applicable to computer systems in general as well as specifically to single-system image cluster systems.